Gemini 3.8 Live: Upgrade for Real-time Voice Models for Voice Agents

Google has introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two new language models designed for real-time audio dialogues and use in voice agents. The models are now available to developers via the Gemini API and Google AI Studio and will be gradually rolled out in Search Live and, in the case of Extended Thinking, also in Google Workspace.

As Google explains in its blog, Gemini 3.8 Live is designed as a cost-effective base model for scalable voice agents. It processes text, image, audio, and video inputs. The context window limit is around 128,000 tokens. According to Google, the model supports 97 languages – including German – and automatically switches between them during a conversation.

The Extended Thinking variant, according to Google, is aimed at complex tasks with multi-step reasoning. It can apparently think in the background while speaking, verbally comment on intermediate results, and process multiple function calls simultaneously.

Asynchronous Function Calls

Technically, Google relies on asynchronous function calls for the Live models, among other things. According to the developer documentation, the mode is standard. This allows the models to call tools in the background while the conversation continues. The scheduling modes SILENT, WHEN_IDLE, and INTERRUPTED can control when and how function results are incorporated into the dialogue.

According to Artificial Analysis, Gemini 3.8 Live Extended Thinking ranks first in the Speech-to-Speech Quality Index with 82.6 points – ahead of OpenAI’s GPT-Live-1 Astra (81.5) and Grok Voice Think Fast 2.0 High (81.3). However, in reasoning, it lags two Qwen models and the StepAudio 3 Realtime model from the Shanghai-registered company StepFun. It leads in agentic performance.

Interestingly, in the Speech Agent Arena, where human users choose the better of two models in a blind test, Extended Thinking is seven places behind Live, and both are behind their predecessor model Gemini 3.1 Flash Live Minimal. They also lag their predecessor in the time to first audio output in this ranking.

Rollout in Google Products and Workspace

The new version of Live is available to developers via the Gemini API and Google AI Studio and will initially be offered to companies as a preview in Gemini Enterprise. For end-users, Gemini 3.8 Live is available in the Live Google Search.

Gemini 3.8 Live Extended Thinking will be available in Gemini Live, among other applications, and for Google AI subscribers in applications such as Docs (for Pro and Ultra only), Gmail, and Keep. This builds on Google’s recently launched AI-powered voice control in Workspace. Companies will receive the feature initially in a private preview. For developers, it is also available via the API and in Google AI Studio.

All audio content generated by Google’s AI products will be watermarked with SynthID. The imperceptible markings are designed to remain machine-readable – a measure that is also relevant against the backdrop of the transparency obligations of the EU AI Act (Article 50) applicable since August 2, 2026.

(rie)